Chromium Code Reviews| Index: runtime/vm/precompiler.cc |
| diff --git a/runtime/vm/precompiler.cc b/runtime/vm/precompiler.cc |
| index f5098bbaa083883da63835398e1010034a56cff7..414ece846ee0738b32bc4dd4d48d1701209c657c 100644 |
| --- a/runtime/vm/precompiler.cc |
| +++ b/runtime/vm/precompiler.cc |
| @@ -4,17 +4,38 @@ |
| #include "vm/precompiler.h" |
| +#include "vm/assembler.h" |
| +#include "vm/ast_printer.h" |
| #include "vm/cha.h" |
| +#include "vm/code_generator.h" |
| #include "vm/code_patcher.h" |
| #include "vm/compiler.h" |
| +#include "vm/constant_propagator.h" |
| +#include "vm/dart_entry.h" |
| +#include "vm/exceptions.h" |
| +#include "vm/flags.h" |
| +#include "vm/flow_graph.h" |
| +#include "vm/flow_graph_allocator.h" |
| +#include "vm/flow_graph_builder.h" |
| +#include "vm/flow_graph_compiler.h" |
| +#include "vm/flow_graph_inliner.h" |
| +#include "vm/flow_graph_optimizer.h" |
| +#include "vm/flow_graph_type_propagator.h" |
| #include "vm/hash_table.h" |
| +#include "vm/il_printer.h" |
| #include "vm/isolate.h" |
| #include "vm/log.h" |
| #include "vm/longjump.h" |
| #include "vm/object.h" |
| #include "vm/object_store.h" |
| +#include "vm/os.h" |
| +#include "vm/parser.h" |
| +#include "vm/regexp_assembler.h" |
| +#include "vm/regexp_parser.h" |
| #include "vm/resolver.h" |
| #include "vm/symbols.h" |
| +#include "vm/tags.h" |
| +#include "vm/timer.h" |
| namespace dart { |
| @@ -29,6 +50,58 @@ DEFINE_FLAG(bool, collect_dynamic_function_names, false, |
| " identify unique targets"); |
| DEFINE_FLAG(bool, print_unique_targets, false, "Print unique dynaic targets"); |
| DEFINE_FLAG(bool, trace_precompiler, false, "Trace precompiler."); |
| +DEFINE_FLAG(int, max_speculative_inlining_attempts, 1, |
| + "Max number of attempts with speculative inlining (precompilation only)"); |
| + |
| +DECLARE_FLAG(bool, allocation_sinking); |
| +DECLARE_FLAG(bool, common_subexpression_elimination); |
| +DECLARE_FLAG(bool, constant_propagation); |
| +DECLARE_FLAG(bool, disassemble); |
| +DECLARE_FLAG(bool, disassemble_optimized); |
| +DECLARE_FLAG(bool, loop_invariant_code_motion); |
| +DECLARE_FLAG(bool, print_flow_graph); |
| +DECLARE_FLAG(bool, print_flow_graph_optimized); |
| +DECLARE_FLAG(bool, range_analysis); |
| +DECLARE_FLAG(bool, trace_compiler); |
| +DECLARE_FLAG(bool, trace_optimizing_compiler); |
| +DECLARE_FLAG(bool, trace_bailout); |
| +DECLARE_FLAG(bool, use_inlining); |
| +DECLARE_FLAG(bool, verify_compiler); |
| +DECLARE_FLAG(bool, precompilation); |
| +DECLARE_FLAG(bool, huge_method_cutoff_in_code_size); |
| +DECLARE_FLAG(bool, load_deferred_eagerly); |
| +DECLARE_FLAG(bool, trace_failed_optimization_attempts); |
| +DECLARE_FLAG(bool, trace_inlining_intervals); |
| +DECLARE_FLAG(bool, trace_irregexp); |
| + |
| + |
| +class PrecompileParsedFunctionHelper : public ValueObject { |
| + public: |
| + PrecompileParsedFunctionHelper(ParsedFunction* parsed_function, |
| + bool optimized) |
| + : parsed_function_(parsed_function), |
| + optimized_(optimized), |
| + thread_(Thread::Current()) { |
| + } |
| + |
| + bool Compile(CompilationPipeline* pipeline); |
| + |
| + private: |
| + ParsedFunction* parsed_function() const { return parsed_function_; } |
| + bool optimized() const { return optimized_; } |
| + Thread* thread() const { return thread_; } |
| + Isolate* isolate() const { return thread_->isolate(); } |
| + |
| + void FinalizeCompilation(Assembler* assembler, |
| + FlowGraphCompiler* graph_compiler, |
| + FlowGraph* flow_graph); |
| + |
| + ParsedFunction* parsed_function_; |
| + const bool optimized_; |
| + Thread* const thread_; |
| + |
| + DISALLOW_COPY_AND_ASSIGN(PrecompileParsedFunctionHelper); |
| +}; |
| static void Jump(const Error& error) { |
| @@ -394,7 +467,7 @@ void Precompiler::ProcessFunction(const Function& function) { |
| ASSERT(!function.is_abstract()); |
| ASSERT(!function.IsRedirectingFactory()); |
| - error_ = Compiler::CompileFunction(thread_, function); |
| + error_ = CompileFunction(thread_, function); |
| if (!error_.IsNull()) { |
| Jump(error_); |
| } |
| @@ -578,7 +651,11 @@ void Precompiler::AddField(const Field& field) { |
| } |
| ASSERT(!Dart::IsRunningPrecompiledCode()); |
| field.SetStaticValue(Instance::Handle(field.SavedInitialStaticValue())); |
| - Compiler::CompileStaticInitializer(field); |
| + const Function& initializer = |
| + Function::Handle(CompileStaticInitializer(field)); |
| + if (!initializer.IsNull()) { |
| + field.SetPrecompiledInitializer(initializer); |
| + } |
| } |
| const Function& function = |
| @@ -589,6 +666,129 @@ void Precompiler::AddField(const Field& field) { |
| } |
| +RawFunction* Precompiler::CompileStaticInitializer(const Field& field) { |
| + ASSERT(field.is_static()); |
| + if (field.HasPrecompiledInitializer()) { |
| + // TODO(rmacnak): Investigate why this happens for _enum_names. |
| + OS::Print("Warning: Ignoring repeated request for initializer for %s\n", |
| + field.ToCString()); |
| + return Function::null(); |
| + } |
| + Thread* thread = Thread::Current(); |
| + StackZone zone(thread); |
| + |
| + ParsedFunction* parsed_function = Parser::ParseStaticFieldInitializer(field); |
| + |
| + parsed_function->AllocateVariables(); |
| + // Non-optimized code generator. |
| + DartCompilationPipeline pipeline; |
| + PrecompileParsedFunctionHelper helper(parsed_function, |
| + /* optimized = */ false); |
| + helper.Compile(&pipeline); |
| + return parsed_function->function().raw(); |
| +} |
| + |
| + |
| +RawObject* Precompiler::EvaluateStaticInitializer(const Field& field) { |
| + ASSERT(field.is_static()); |
| + // The VM sets the field's value to transiton_sentinel prior to |
| + // evaluating the initializer value. |
| + ASSERT(field.StaticValue() == Object::transition_sentinel().raw()); |
| + LongJumpScope jump; |
| + if (setjmp(*jump.Set()) == 0) { |
| + // Under precompilation, the initializer may have already been compiled, in |
| + // which case use it. Under lazy compilation or early in precompilation, the |
| + // initializer has not yet been created, so create it now, but don't bother |
| + // remembering it because it won't be used again. |
| + Function& initializer = Function::Handle(); |
| + if (!field.HasPrecompiledInitializer()) { |
| + initializer = CompileStaticInitializer(field); |
| + Code::Handle(initializer.unoptimized_code()).set_var_descriptors( |
| + Object::empty_var_descriptors()); |
| + } else { |
| + initializer ^= field.PrecompiledInitializer(); |
| + } |
| + // Invoke the function to evaluate the expression. |
| + return DartEntry::InvokeFunction(initializer, Object::empty_array()); |
| + } else { |
| + Thread* const thread = Thread::Current(); |
| + Isolate* const isolate = thread->isolate(); |
| + StackZone zone(thread); |
| + const Error& error = |
| + Error::Handle(thread->zone(), isolate->object_store()->sticky_error()); |
| + isolate->object_store()->clear_sticky_error(); |
| + return error.raw(); |
| + } |
| + UNREACHABLE(); |
| + return Object::null(); |
| +} |
| + |
| + |
| +RawObject* Precompiler::ExecuteOnce(SequenceNode* fragment) { |
| + LongJumpScope jump; |
| + if (setjmp(*jump.Set()) == 0) { |
| + Thread* const thread = Thread::Current(); |
| + if (FLAG_trace_compiler) { |
| + THR_Print("compiling expression: "); |
| + AstPrinter::PrintNode(fragment); |
| + } |
| + |
| + // Create a dummy function object for the code generator. |
| + // The function needs to be associated with a named Class: the interface |
| + // Function fits the bill. |
| + const char* kEvalConst = "eval_const"; |
| + const Function& func = Function::ZoneHandle(Function::New( |
| + String::Handle(Symbols::New(kEvalConst)), |
| + RawFunction::kRegularFunction, |
| + true, // static function |
| + false, // not const function |
| + false, // not abstract |
| + false, // not external |
| + false, // not native |
| + Class::Handle(Type::Handle(Type::Function()).type_class()), |
| + fragment->token_pos())); |
| + |
| + func.set_result_type(Object::dynamic_type()); |
| + func.set_num_fixed_parameters(0); |
| + func.SetNumOptionalParameters(0, true); |
| + // Manually generated AST, do not recompile. |
| + func.SetIsOptimizable(false); |
| + func.set_is_debuggable(false); |
| + |
| + // We compile the function here, even though InvokeFunction() below |
| + // would compile func automatically. We are checking fewer invariants |
| + // here. |
| + ParsedFunction* parsed_function = new ParsedFunction(thread, func); |
| + parsed_function->SetNodeSequence(fragment); |
| + fragment->scope()->AddVariable(parsed_function->EnsureExpressionTemp()); |
| + fragment->scope()->AddVariable( |
| + parsed_function->current_context_var()); |
| + parsed_function->AllocateVariables(); |
| + |
| + // Non-optimized code generator. |
| + DartCompilationPipeline pipeline; |
| + PrecompileParsedFunctionHelper helper(parsed_function, |
| + /* optimized = */ false); |
| + helper.Compile(&pipeline); |
| + Code::Handle(func.unoptimized_code()).set_var_descriptors( |
| + Object::empty_var_descriptors()); |
| + |
| + const Object& result = PassiveObject::Handle( |
| + DartEntry::InvokeFunction(func, Object::empty_array())); |
| + return result.raw(); |
| + } else { |
| + Thread* const thread = Thread::Current(); |
| + Isolate* const isolate = thread->isolate(); |
| + const Object& result = |
| + PassiveObject::Handle(isolate->object_store()->sticky_error()); |
| + isolate->object_store()->clear_sticky_error(); |
| + return result.raw(); |
| + } |
| + UNREACHABLE(); |
| + return Object::null(); |
| +} |
| + |
| + |
| void Precompiler::AddFunction(const Function& function) { |
| if (enqueued_functions_.Lookup(&function) != NULL) return; |
| @@ -1272,4 +1472,650 @@ void Precompiler::ResetPrecompilerState() { |
| } |
| } |
| + |
| +void PrecompileParsedFunctionHelper::FinalizeCompilation( |
| + Assembler* assembler, |
| + FlowGraphCompiler* graph_compiler, |
| + FlowGraph* flow_graph) { |
| + const Function& function = parsed_function()->function(); |
| + Zone* const zone = thread()->zone(); |
| + |
| + CSTAT_TIMER_SCOPE(thread(), codefinalizer_timer); |
| + // CreateDeoptInfo uses the object pool and needs to be done before |
| + // FinalizeCode. |
| + const Array& deopt_info_array = |
| + Array::Handle(zone, graph_compiler->CreateDeoptInfo(assembler)); |
| + INC_STAT(thread(), total_code_size, |
| + deopt_info_array.Length() * sizeof(uword)); |
| + // Allocates instruction object. Since this occurs only at safepoint, |
| + // there can be no concurrent access to the instruction page. |
| + const Code& code = Code::Handle( |
| + Code::FinalizeCode(function, assembler, optimized())); |
| + code.set_is_optimized(optimized()); |
| + code.set_owner(function); |
| + if (!function.IsOptimizable()) { |
| + // A function with huge unoptimized code can become non-optimizable |
| + // after generating unoptimized code. |
| + function.set_usage_counter(INT_MIN); |
| + } |
| + |
| + const Array& intervals = graph_compiler->inlined_code_intervals(); |
| + INC_STAT(thread(), total_code_size, |
| + intervals.Length() * sizeof(uword)); |
| + code.SetInlinedIntervals(intervals); |
| + |
| + const Array& inlined_id_array = |
| + Array::Handle(zone, graph_compiler->InliningIdToFunction()); |
| + INC_STAT(thread(), total_code_size, |
| + inlined_id_array.Length() * sizeof(uword)); |
| + code.SetInlinedIdToFunction(inlined_id_array); |
| + |
| + const Array& caller_inlining_id_map_array = |
| + Array::Handle(zone, graph_compiler->CallerInliningIdMap()); |
| + INC_STAT(thread(), total_code_size, |
| + caller_inlining_id_map_array.Length() * sizeof(uword)); |
| + code.SetInlinedCallerIdMap(caller_inlining_id_map_array); |
| + |
| + graph_compiler->FinalizePcDescriptors(code); |
| + code.set_deopt_info_array(deopt_info_array); |
| + |
| + graph_compiler->FinalizeStackmaps(code); |
| + graph_compiler->FinalizeVarDescriptors(code); |
| + graph_compiler->FinalizeExceptionHandlers(code); |
| + graph_compiler->FinalizeStaticCallTargetsTable(code); |
| + |
| + if (optimized()) { |
| + // Installs code while at safepoint. |
| + ASSERT(thread()->IsMutatorThread()); |
| + function.InstallOptimizedCode(code, /* is_osr = */ false); |
| + } else { // not optimized. |
| + function.set_unoptimized_code(code); |
| + function.AttachCode(code); |
| + } |
| + ASSERT(!parsed_function()->HasDeferredPrefixes()); |
| + ASSERT(FLAG_load_deferred_eagerly); |
| +} |
| + |
| + |
| +// Return false if bailed out. |
| +// If optimized_result_code is not NULL then it is caller's responsibility |
| +// to install code. |
| +bool PrecompileParsedFunctionHelper::Compile(CompilationPipeline* pipeline) { |
| + ASSERT(FLAG_precompilation); |
| + const Function& function = parsed_function()->function(); |
| + if (optimized() && !function.IsOptimizable()) { |
| + return false; |
| + } |
| + bool is_compiled = false; |
| + Zone* const zone = thread()->zone(); |
| + TimelineStream* compiler_timeline = isolate()->GetCompilerStream(); |
| + CSTAT_TIMER_SCOPE(thread(), codegen_timer); |
| + HANDLESCOPE(thread()); |
| + |
| + // We may reattempt compilation if the function needs to be assembled using |
| + // far branches on ARM and MIPS. In the else branch of the setjmp call, |
| + // done is set to false, and use_far_branches is set to true if there is a |
| + // longjmp from the ARM or MIPS assemblers. In all other paths through this |
| + // while loop, done is set to true. use_far_branches is always false on ia32 |
| + // and x64. |
| + bool done = false; |
| + // volatile because the variable may be clobbered by a longjmp. |
| + volatile bool use_far_branches = false; |
| + volatile bool use_speculative_inlining = |
| + FLAG_max_speculative_inlining_attempts > 0; |
| + GrowableArray<intptr_t> inlining_black_list; |
| + |
| + while (!done) { |
| + const intptr_t prev_deopt_id = thread()->deopt_id(); |
| + thread()->set_deopt_id(0); |
| + LongJumpScope jump; |
| + const intptr_t val = setjmp(*jump.Set()); |
| + if (val == 0) { |
| + FlowGraph* flow_graph = NULL; |
| + |
| + // Class hierarchy analysis is registered with the isolate in the |
| + // constructor and unregisters itself upon destruction. |
| + CHA cha(thread()); |
| + |
| + // TimerScope needs an isolate to be properly terminated in case of a |
| + // LongJump. |
| + { |
| + CSTAT_TIMER_SCOPE(thread(), graphbuilder_timer); |
| + ZoneGrowableArray<const ICData*>* ic_data_array = |
| + new(zone) ZoneGrowableArray<const ICData*>(); |
| + TimelineDurationScope tds(thread(), |
| + compiler_timeline, |
| + "BuildFlowGraph"); |
| + flow_graph = pipeline->BuildFlowGraph(zone, |
| + parsed_function(), |
| + *ic_data_array, |
| + Compiler::kNoOSRDeoptId); |
| + } |
| + |
| + const bool print_flow_graph = |
| + (FLAG_print_flow_graph || |
| + (optimized() && FLAG_print_flow_graph_optimized)) && |
| + FlowGraphPrinter::ShouldPrint(function); |
| + |
| + if (print_flow_graph) { |
| + FlowGraphPrinter::PrintGraph("Before Optimizations", flow_graph); |
| + } |
| + |
| + if (optimized()) { |
| + TimelineDurationScope tds(thread(), |
| + compiler_timeline, |
| + "ComputeSSA"); |
| + CSTAT_TIMER_SCOPE(thread(), ssa_timer); |
| + // Transform to SSA (virtual register 0 and no inlining arguments). |
| + flow_graph->ComputeSSA(0, NULL); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + if (print_flow_graph) { |
| + FlowGraphPrinter::PrintGraph("After SSA", flow_graph); |
| + } |
| + } |
| + |
| + // Maps inline_id_to_function[inline_id] -> function. Top scope |
| + // function has inline_id 0. The map is populated by the inliner. |
| + GrowableArray<const Function*> inline_id_to_function; |
| + // For a given inlining-id(index) specifies the caller's inlining-id. |
| + GrowableArray<intptr_t> caller_inline_id; |
| + // Collect all instance fields that are loaded in the graph and |
| + // have non-generic type feedback attached to them that can |
| + // potentially affect optimizations. |
| + if (optimized()) { |
| + TimelineDurationScope tds(thread(), |
| + compiler_timeline, |
| + "OptimizationPasses"); |
| + inline_id_to_function.Add(&function); |
| + // Top scope function has no caller (-1). |
| + caller_inline_id.Add(-1); |
| + CSTAT_TIMER_SCOPE(thread(), graphoptimizer_timer); |
| + |
| + FlowGraphOptimizer optimizer(flow_graph, |
| + use_speculative_inlining, |
| + &inlining_black_list); |
| + optimizer.PopulateWithICData(); |
| + |
| + optimizer.ApplyClassIds(); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + FlowGraphTypePropagator::Propagate(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + optimizer.ApplyICData(); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + // Optimize (a << b) & c patterns, merge operations. |
| + // Run early in order to have more opportunity to optimize left shifts. |
| + optimizer.TryOptimizePatterns(); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + FlowGraphInliner::SetInliningId(flow_graph, 0); |
| + |
| + // Inlining (mutates the flow graph) |
| + if (FLAG_use_inlining) { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "Inlining"); |
| + CSTAT_TIMER_SCOPE(thread(), graphinliner_timer); |
| + // Propagate types to create more inlining opportunities. |
| + FlowGraphTypePropagator::Propagate(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + // Use propagated class-ids to create more inlining opportunities. |
| + optimizer.ApplyClassIds(); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + FlowGraphInliner inliner(flow_graph, |
| + &inline_id_to_function, |
| + &caller_inline_id, |
| + use_speculative_inlining, |
| + &inlining_black_list); |
| + inliner.Inline(); |
| + // Use lists are maintained and validated by the inliner. |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + } |
| + |
| + // Propagate types and eliminate more type tests. |
| + FlowGraphTypePropagator::Propagate(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "ApplyClassIds"); |
| + // Use propagated class-ids to optimize further. |
| + optimizer.ApplyClassIds(); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + } |
| + |
| + // Propagate types for potentially newly added instructions by |
| + // ApplyClassIds(). Must occur before canonicalization. |
| + FlowGraphTypePropagator::Propagate(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + // Do optimizations that depend on the propagated type information. |
| + if (optimizer.Canonicalize()) { |
| + // Invoke Canonicalize twice in order to fully canonicalize patterns |
| + // like "if (a & const == 0) { }". |
| + optimizer.Canonicalize(); |
| + } |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "BranchSimplifier"); |
| + BranchSimplifier::Simplify(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + IfConverter::Simplify(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + } |
| + |
| + if (FLAG_constant_propagation) { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "ConstantPropagation"); |
| + ConstantPropagator::Optimize(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + // A canonicalization pass to remove e.g. smi checks on smi constants. |
| + optimizer.Canonicalize(); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + // Canonicalization introduced more opportunities for constant |
| + // propagation. |
| + ConstantPropagator::Optimize(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + } |
| + |
| + // Optimistically convert loop phis that have a single non-smi input |
| + // coming from the loop pre-header into smi-phis. |
| + if (FLAG_loop_invariant_code_motion) { |
| + LICM licm(flow_graph); |
| + licm.OptimisticallySpecializeSmiPhis(); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + } |
| + |
| + // Propagate types and eliminate even more type tests. |
| + // Recompute types after constant propagation to infer more precise |
| + // types for uses that were previously reached by now eliminated phis. |
| + FlowGraphTypePropagator::Propagate(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "SelectRepresentations"); |
|
rmacnak
2016/02/04 00:36:39
Maybe we should skip these two.
Florian Schneider
2016/02/05 01:55:52
I'd rather leave them in for now. They may actuall
rmacnak
2016/02/08 22:43:23
Acknowledged.
|
| + // Where beneficial convert Smi operations into Int32 operations. |
| + // Only meanigful for 32bit platforms right now. |
| + optimizer.WidenSmiToInt32(); |
| + |
| + // Unbox doubles. Performed after constant propagation to minimize |
| + // interference from phis merging double values and tagged |
| + // values coming from dead paths. |
| + optimizer.SelectRepresentations(); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + } |
| + |
| + { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "CommonSubexpressionElinination"); |
| + if (FLAG_common_subexpression_elimination || |
| + FLAG_loop_invariant_code_motion) { |
| + flow_graph->ComputeBlockEffects(); |
| + } |
| + |
| + if (FLAG_common_subexpression_elimination) { |
| + if (DominatorBasedCSE::Optimize(flow_graph)) { |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + optimizer.Canonicalize(); |
| + // Do another round of CSE to take secondary effects into account: |
| + // e.g. when eliminating dependent loads (a.x[0] + a.x[0]) |
| + // TODO(fschneider): Change to a one-pass optimization pass. |
| + if (DominatorBasedCSE::Optimize(flow_graph)) { |
| + optimizer.Canonicalize(); |
| + } |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + } |
| + } |
| + |
| + // Run loop-invariant code motion right after load elimination since |
| + // it depends on the numbering of loads from the previous |
| + // load-elimination. |
| + if (FLAG_loop_invariant_code_motion) { |
| + LICM licm(flow_graph); |
| + licm.Optimize(); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + } |
| + flow_graph->RemoveRedefinitions(); |
| + } |
| + |
| + // Optimize (a << b) & c patterns, merge operations. |
| + // Run after CSE in order to have more opportunity to merge |
| + // instructions that have same inputs. |
| + optimizer.TryOptimizePatterns(); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "DeadStoreElimination"); |
| + DeadStoreElimination::Optimize(flow_graph); |
| + } |
| + |
| + if (FLAG_range_analysis) { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "RangeAnalysis"); |
| + // Propagate types after store-load-forwarding. Some phis may have |
| + // become smi phis that can be processed by range analysis. |
| + FlowGraphTypePropagator::Propagate(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + // We have to perform range analysis after LICM because it |
| + // optimistically moves CheckSmi through phis into loop preheaders |
| + // making some phis smi. |
| + optimizer.InferIntRanges(); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + } |
| + |
| + if (FLAG_constant_propagation) { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "ConstantPropagator::OptimizeBranches"); |
| + // Constant propagation can use information from range analysis to |
| + // find unreachable branch targets and eliminate branches that have |
| + // the same true- and false-target. |
| + ConstantPropagator::OptimizeBranches(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + } |
| + |
| + // Recompute types after code movement was done to ensure correct |
| + // reaching types for hoisted values. |
| + FlowGraphTypePropagator::Propagate(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "TryCatchAnalyzer::Optimize"); |
| + // Optimize try-blocks. |
| + TryCatchAnalyzer::Optimize(flow_graph); |
| + } |
| + |
| + // Detach environments from the instructions that can't deoptimize. |
| + // Do it before we attempt to perform allocation sinking to minimize |
| + // amount of materializations it has to perform. |
| + optimizer.EliminateEnvironments(); |
| + |
| + { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "EliminateDeadPhis"); |
| + DeadCodeElimination::EliminateDeadPhis(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + } |
| + |
| + if (optimizer.Canonicalize()) { |
| + optimizer.Canonicalize(); |
| + } |
| + |
| + // Attempt to sink allocations of temporary non-escaping objects to |
| + // the deoptimization path. |
| + AllocationSinking* sinking = NULL; |
| + if (FLAG_allocation_sinking && |
| + (flow_graph->graph_entry()->SuccessorCount() == 1)) { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "AllocationSinking::Optimize"); |
| + // TODO(fschneider): Support allocation sinking with try-catch. |
| + sinking = new AllocationSinking(flow_graph); |
| + sinking->Optimize(); |
| + } |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + DeadCodeElimination::EliminateDeadPhis(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + FlowGraphTypePropagator::Propagate(flow_graph); |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "SelectRepresentations"); |
| + // Ensure that all phis inserted by optimization passes have |
| + // consistent representations. |
| + optimizer.SelectRepresentations(); |
| + } |
| + |
| + if (optimizer.Canonicalize()) { |
| + // To fully remove redundant boxing (e.g. BoxDouble used only in |
| + // environments and UnboxDouble instructions) instruction we |
| + // first need to replace all their uses and then fold them away. |
| + // For now we just repeat Canonicalize twice to do that. |
| + // TODO(vegorov): implement a separate representation folding pass. |
| + optimizer.Canonicalize(); |
| + } |
| + DEBUG_ASSERT(flow_graph->VerifyUseLists()); |
| + |
| + if (sinking != NULL) { |
| + TimelineDurationScope tds2( |
| + thread(), |
| + compiler_timeline, |
| + "AllocationSinking::DetachMaterializations"); |
| + // Remove all MaterializeObject instructions inserted by allocation |
| + // sinking from the flow graph and let them float on the side |
| + // referenced only from environments. Register allocator will consider |
| + // them as part of a deoptimization environment. |
| + sinking->DetachMaterializations(); |
| + } |
| + |
| + // Compute and store graph informations (call & instruction counts) |
| + // to be later used by the inliner. |
| + FlowGraphInliner::CollectGraphInfo(flow_graph, true); |
| + |
| + { |
| + TimelineDurationScope tds2(thread(), |
| + compiler_timeline, |
| + "AllocateRegisters"); |
| + // Perform register allocation on the SSA graph. |
| + FlowGraphAllocator allocator(*flow_graph); |
| + allocator.AllocateRegisters(); |
| + } |
| + |
| + if (print_flow_graph) { |
| + FlowGraphPrinter::PrintGraph("After Optimizations", flow_graph); |
| + } |
| + } |
| + |
| + ASSERT(inline_id_to_function.length() == caller_inline_id.length()); |
| + Assembler assembler(use_far_branches); |
| + FlowGraphCompiler graph_compiler(&assembler, flow_graph, |
| + *parsed_function(), optimized(), |
| + inline_id_to_function, |
| + caller_inline_id); |
| + { |
| + CSTAT_TIMER_SCOPE(thread(), graphcompiler_timer); |
| + TimelineDurationScope tds(thread(), |
| + compiler_timeline, |
| + "CompileGraph"); |
| + graph_compiler.CompileGraph(); |
| + pipeline->FinalizeCompilation(); |
| + } |
| + { |
| + TimelineDurationScope tds(thread(), |
| + compiler_timeline, |
| + "FinalizeCompilation"); |
| + ASSERT(thread()->IsMutatorThread()); |
| + FinalizeCompilation(&assembler, &graph_compiler, flow_graph); |
| + } |
| + // Mark that this isolate now has compiled code. |
| + isolate()->set_has_compiled_code(true); |
| + // Exit the loop and the function with the correct result value. |
| + is_compiled = true; |
| + done = true; |
| + } else { |
| + // We bailed out or we encountered an error. |
| + const Error& error = Error::Handle( |
| + isolate()->object_store()->sticky_error()); |
| + |
| + if (error.raw() == Object::branch_offset_error().raw()) { |
| + // Compilation failed due to an out of range branch offset in the |
| + // assembler. We try again (done = false) with far branches enabled. |
| + done = false; |
| + ASSERT(!use_far_branches); |
| + use_far_branches = true; |
| + } else if (error.raw() == Object::speculative_inlining_error().raw()) { |
| + // The return value of setjmp is the deopt id of the check instruction |
| + // that caused the bailout. |
| + done = false; |
| +#if defined(DEBUG) |
| + ASSERT(use_speculative_inlining); |
| + for (intptr_t i = 0; i < inlining_black_list.length(); ++i) { |
| + ASSERT(inlining_black_list[i] != val); |
| + } |
| +#endif |
| + inlining_black_list.Add(val); |
| + const intptr_t max_attempts = FLAG_max_speculative_inlining_attempts; |
| + if (inlining_black_list.length() >= max_attempts) { |
| + use_speculative_inlining = false; |
| + if (FLAG_trace_compiler || FLAG_trace_optimizing_compiler) { |
| + THR_Print("Disabled speculative inlining after %" Pd " attempts.\n", |
| + inlining_black_list.length()); |
| + } |
| + } |
| + } else { |
| + // If the error isn't due to an out of range branch offset, we don't |
| + // try again (done = true), and indicate that we did not finish |
| + // compiling (is_compiled = false). |
| + if (FLAG_trace_bailout) { |
| + THR_Print("%s\n", error.ToErrorCString()); |
| + } |
| + done = true; |
| + } |
| + |
| + // Clear the error if it was not a real error, but just a bailout. |
| + if (error.IsLanguageError() && |
| + (LanguageError::Cast(error).kind() == Report::kBailout)) { |
| + isolate()->object_store()->clear_sticky_error(); |
| + } |
| + is_compiled = false; |
| + } |
| + // Reset global isolate state. |
| + thread()->set_deopt_id(prev_deopt_id); |
| + } |
| + return is_compiled; |
| +} |
| + |
| + |
| +static RawError* PrecompileFunctionHelper(CompilationPipeline* pipeline, |
| + const Function& function, |
| + bool optimized) { |
| + // Check that we optimize, except if the function is not optimizable. |
| + ASSERT(FLAG_precompilation); |
| + ASSERT(!function.IsOptimizable() || optimized); |
| + ASSERT(!function.HasCode()); |
| + LongJumpScope jump; |
| + if (setjmp(*jump.Set()) == 0) { |
| + Thread* const thread = Thread::Current(); |
| + Isolate* const isolate = thread->isolate(); |
| + StackZone stack_zone(thread); |
| + Zone* const zone = stack_zone.GetZone(); |
| + const bool trace_compiler = |
| + FLAG_trace_compiler || |
| + (FLAG_trace_optimizing_compiler && optimized); |
| + Timer per_compile_timer(trace_compiler, "Compilation time"); |
| + per_compile_timer.Start(); |
| + |
| + ParsedFunction* parsed_function = new(zone) ParsedFunction( |
| + thread, Function::ZoneHandle(zone, function.raw())); |
| + if (trace_compiler) { |
| + THR_Print( |
| + "Precompiling %sfunction: '%s' @ token %" Pd ", size %" Pd "\n", |
| + (optimized ? "optimized " : ""), |
| + function.ToFullyQualifiedCString(), |
| + function.token_pos().Pos(), |
| + (function.end_token_pos().Pos() - function.token_pos().Pos())); |
| + } |
| + INC_STAT(thread, num_functions_compiled, 1); |
| + if (optimized) { |
| + INC_STAT(thread, num_functions_optimized, 1); |
| + } |
| + { |
| + HANDLESCOPE(thread); |
| + const int64_t num_tokens_before = STAT_VALUE(thread, num_tokens_consumed); |
| + pipeline->ParseFunction(parsed_function); |
| + const int64_t num_tokens_after = STAT_VALUE(thread, num_tokens_consumed); |
| + INC_STAT(thread, |
| + num_func_tokens_compiled, |
| + num_tokens_after - num_tokens_before); |
| + } |
| + |
| + PrecompileParsedFunctionHelper helper(parsed_function, optimized); |
| + const bool success = helper.Compile(pipeline); |
| + if (!success) { |
| + // Encountered error. |
| + Error& error = Error::Handle(); |
| + // We got an error during compilation. |
| + error = isolate->object_store()->sticky_error(); |
| + isolate->object_store()->clear_sticky_error(); |
| + ASSERT(error.IsLanguageError() && |
| + LanguageError::Cast(error).kind() != Report::kBailout); |
| + return error.raw(); |
| + } |
| + |
| + per_compile_timer.Stop(); |
| + |
| + if (trace_compiler && success) { |
| + THR_Print("--> '%s' entry: %#" Px " size: %" Pd " time: %" Pd64 " us\n", |
| + function.ToFullyQualifiedCString(), |
| + Code::Handle(function.CurrentCode()).EntryPoint(), |
| + Code::Handle(function.CurrentCode()).Size(), |
| + per_compile_timer.TotalElapsedTime()); |
| + } |
| + |
| + if (FLAG_disassemble && FlowGraphPrinter::ShouldPrint(function)) { |
| + Compiler::DisassembleCode(function, optimized); |
| + } else if (FLAG_disassemble_optimized && |
| + optimized && |
| + FlowGraphPrinter::ShouldPrint(function)) { |
| + // TODO(fschneider): Print unoptimized code along with the optimized code. |
| + THR_Print("*** BEGIN CODE\n"); |
| + Compiler::DisassembleCode(function, true); |
| + THR_Print("*** END CODE\n"); |
| + } |
| + return Error::null(); |
| + } else { |
| + Thread* const thread = Thread::Current(); |
| + Isolate* const isolate = thread->isolate(); |
| + StackZone stack_zone(thread); |
| + Error& error = Error::Handle(); |
| + // We got an error during compilation. |
| + error = isolate->object_store()->sticky_error(); |
| + isolate->object_store()->clear_sticky_error(); |
| + // Precompilation may encounter compile-time errors. |
| + // Do not attempt to optimize functions that can cause errors. |
| + function.set_is_optimizable(false); |
| + return error.raw(); |
| + } |
| + UNREACHABLE(); |
| + return Error::null(); |
| +} |
| + |
| + |
| +RawError* Precompiler::CompileFunction(Thread* thread, |
| + const Function& function) { |
| + VMTagScope tagScope(thread, VMTag::kCompileUnoptimizedTagId); |
| + TIMELINE_FUNCTION_COMPILATION_DURATION(thread, "Function", function); |
| + |
| + CompilationPipeline* pipeline = |
| + CompilationPipeline::New(thread->zone(), function); |
| + |
| + ASSERT(FLAG_precompilation); |
| + const bool optimized = function.IsOptimizable(); // False for natives. |
| + return PrecompileFunctionHelper(pipeline, function, optimized); |
| +} |
| + |
| } // namespace dart |